Skip to content

[ET-VK] Using shared variable to store calculated output pose to free up registers and improve performance. #7503

New issue

Have a question about this project? Sign up for a free GitHub account to open an issue and contact its maintainers and the community.

Sign up for GitHub

By clicking “Sign up for GitHub”, you agree to our terms of service and privacy statement. We’ll occasionally send you account related emails.

Already on GitHub? Sign in to your account

Merged
trivedivivek merged 6 commits into from
Jan 4, 2025
Merged

[ET-VK] Using shared variable to store calculated output pose to free up registers and improve performance. #7503

Changes from all commits
Commits
Show all changes
6 commits
Select commit Hold shift + click to select a range
7c31a5d
[ET-VK] Reduced int precision for all int storage in conv pw op to im…
trivedivivek Jan 3, 2025
f139e39
[ET-VK] Minor fix to conv 2d op using wg_size from create_conv2d_glob…
trivedivivek Jan 3, 2025
5b1c980
[ET-VK] Modify conv 2d pw op shader and dispatch settings to linearly…
trivedivivek Jan 3, 2025
c913e17
[ET-VK] Using vec2 to store output positions to reudce shader registe…
trivedivivek Jan 3, 2025
c805499
[ET-VK] Using shared variable to store calculated output pose to free…
trivedivivek Jan 3, 2025
5b34f29
Merge branch 'main' into gh/trivedivivek/33/orig
trivedivivek Jan 4, 2025
File filter

Filter by extension

Filter by extension
Conversations
Failed to load comments.
Loading
Jump to
Jump to file
Failed to load files.
Loading
Diff view
Diff view
12 changes: 9 additions & 3 deletions backends/vulkan/runtime/graph/ops/glsl/conv2d_pw.glsl
View file
Open in desktop
Original file line number Diff line number Diff line change
Expand Up @@ -34,13 +34,17 @@ layout(local_size_x_id = 0, local_size_y_id = 1, local_size_z_id = 2) in;

#extension GL_EXT_shader_explicit_arithmetic_types_int16 : require

// shared memory to hold calculated positions, this would reduce register usage thus improving performance.
shared u16vec2 pos_shared[gl_WorkGroupSize.x * gl_WorkGroupSize.y * gl_WorkGroupSize.z * TILE_SIZE * TILE_SIZE];

/*
* Computes a 2D pointwise convolution of an NxN output tile. Calculating an
* output tile for pointwise convolution is more efficient because the kernel
* size is only 1x1, making it easier to re-use loaded texels from t_kernel.
*/
void main() {
const uint16_t out_limits_y_scaled = uint16_t((out_limits.y + TILE_SIZE - 1) / TILE_SIZE);
const uint shared_mem_stride = gl_WorkGroupSize.x * gl_WorkGroupSize.y * gl_WorkGroupSize.z;

const u16vec3 gpos = u16vec3(
gl_GlobalInvocationID.x / (out_limits_y_scaled * out_limits.z),
Expand All @@ -58,6 +62,7 @@ void main() {
for (int x = 0; x < TILE_SIZE; ++x) {
pos[i] = u16vec2(
gpos.x * TILE_SIZE + x, gpos.y * TILE_SIZE + y);
pos_shared[(shared_mem_stride * i) + gl_LocalInvocationIndex] = pos[i];
i++;
}
}
Expand All @@ -73,7 +78,7 @@ void main() {
// the top-left element is in a region added by padding.
u16vec2 ipos[TILE_SIZE * TILE_SIZE];
for (int i = 0; i < TILE_SIZE * TILE_SIZE; ++i) {
ipos[i] = pos[i].xy * u16vec2(stride) - u16vec2(padding);
ipos[i] = pos[i] * u16vec2(stride) - u16vec2(padding);
}

vec4 sum[TILE_SIZE * TILE_SIZE];
Expand Down Expand Up @@ -138,8 +143,9 @@ void main() {
}

for (int i = 0; i < TILE_SIZE * TILE_SIZE; ++i) {
if (all(lessThan(u16vec3(pos[i], gpos.z), out_limits))) {
imageStore(t_out, u16vec3(pos[i], gpos.z), op(sum[i], out_min, out_max));
const u16vec2 pos = pos_shared[(shared_mem_stride * i) + gl_LocalInvocationIndex];
if (all(lessThan(u16vec3(pos, gpos.z), out_limits))) {
imageStore(t_out, u16vec3(pos, gpos.z), op(sum[i], out_min, out_max));
}
}
}